2 * This file contains job initialization and setup functions.
11 #include <sys/types.h>
13 #ifdef CONFIG_VALGRIND_DEV
14 #include <valgrind/drd.h>
16 #define DRD_IGNORE_VAR(x) do { } while (0)
20 #ifndef FIO_NO_HAVE_SHM_H
32 #include "steadystate.h"
35 #include "oslib/asprintf.h"
36 #include "oslib/getopt.h"
37 #include "oslib/strcasestr.h"
41 #include "lib/memcpy.h"
43 const char fio_version_string[] = FIO_VERSION;
45 #define FIO_RANDSEED (0xb1899bedUL)
47 static char **ini_file;
48 static bool dump_cmdline;
49 static bool parse_only;
50 static bool merge_blktrace_only;
52 static struct thread_data def_thread;
53 struct thread_segment segments[REAL_MAX_SEG];
54 static char **job_sections;
55 static int nr_job_sections;
57 bool exitall_on_terminate = false;
58 int output_format = FIO_OUTPUT_NORMAL;
59 int eta_print = FIO_ETA_AUTO;
60 unsigned int eta_interval_msec = 1000;
64 char *exec_profile = NULL;
65 int warnings_fatal = 0;
66 int terse_version = 3;
67 bool is_backend = false;
68 bool is_local_backend = false;
70 bool log_syslog = false;
72 bool write_bw_log = false;
73 bool read_only = false;
74 int status_interval = 0;
76 char *trigger_file = NULL;
77 long long trigger_timeout = 0;
78 char *trigger_cmd = NULL;
79 char *trigger_remote_cmd = NULL;
81 char *aux_path = NULL;
83 static int prev_group_jobs;
85 unsigned long fio_debug = 0;
86 unsigned int fio_debug_jobno = -1;
87 unsigned int *fio_debug_jobp = NULL;
88 unsigned int *fio_warned = NULL;
90 static char cmd_optstr[256];
93 #define FIO_CLIENT_FLAG (1 << 16)
96 * Command line options. These will contain the above, plus a few
97 * extra that only pertain to fio itself and not jobs.
99 static struct option l_opts[FIO_NR_OPTIONS] = {
101 .name = (char *) "output",
102 .has_arg = required_argument,
103 .val = 'o' | FIO_CLIENT_FLAG,
106 .name = (char *) "latency-log",
107 .has_arg = required_argument,
108 .val = 'l' | FIO_CLIENT_FLAG,
111 .name = (char *) "bandwidth-log",
112 .has_arg = no_argument,
113 .val = 'b' | FIO_CLIENT_FLAG,
116 .name = (char *) "minimal",
117 .has_arg = no_argument,
118 .val = 'm' | FIO_CLIENT_FLAG,
121 .name = (char *) "output-format",
122 .has_arg = required_argument,
123 .val = 'F' | FIO_CLIENT_FLAG,
126 .name = (char *) "append-terse",
127 .has_arg = optional_argument,
131 .name = (char *) "version",
132 .has_arg = no_argument,
133 .val = 'v' | FIO_CLIENT_FLAG,
136 .name = (char *) "help",
137 .has_arg = no_argument,
138 .val = 'h' | FIO_CLIENT_FLAG,
141 .name = (char *) "cmdhelp",
142 .has_arg = optional_argument,
143 .val = 'c' | FIO_CLIENT_FLAG,
146 .name = (char *) "enghelp",
147 .has_arg = optional_argument,
148 .val = 'i' | FIO_CLIENT_FLAG,
151 .name = (char *) "showcmd",
152 .has_arg = no_argument,
153 .val = 's' | FIO_CLIENT_FLAG,
156 .name = (char *) "readonly",
157 .has_arg = no_argument,
158 .val = 'r' | FIO_CLIENT_FLAG,
161 .name = (char *) "eta",
162 .has_arg = required_argument,
163 .val = 'e' | FIO_CLIENT_FLAG,
166 .name = (char *) "eta-interval",
167 .has_arg = required_argument,
168 .val = 'O' | FIO_CLIENT_FLAG,
171 .name = (char *) "eta-newline",
172 .has_arg = required_argument,
173 .val = 'E' | FIO_CLIENT_FLAG,
176 .name = (char *) "debug",
177 .has_arg = required_argument,
178 .val = 'd' | FIO_CLIENT_FLAG,
181 .name = (char *) "parse-only",
182 .has_arg = no_argument,
183 .val = 'P' | FIO_CLIENT_FLAG,
186 .name = (char *) "section",
187 .has_arg = required_argument,
188 .val = 'x' | FIO_CLIENT_FLAG,
192 .name = (char *) "inflate-log",
193 .has_arg = required_argument,
194 .val = 'X' | FIO_CLIENT_FLAG,
198 .name = (char *) "alloc-size",
199 .has_arg = required_argument,
200 .val = 'a' | FIO_CLIENT_FLAG,
203 .name = (char *) "profile",
204 .has_arg = required_argument,
205 .val = 'p' | FIO_CLIENT_FLAG,
208 .name = (char *) "warnings-fatal",
209 .has_arg = no_argument,
210 .val = 'w' | FIO_CLIENT_FLAG,
213 .name = (char *) "max-jobs",
214 .has_arg = required_argument,
215 .val = 'j' | FIO_CLIENT_FLAG,
218 .name = (char *) "terse-version",
219 .has_arg = required_argument,
220 .val = 'V' | FIO_CLIENT_FLAG,
223 .name = (char *) "server",
224 .has_arg = optional_argument,
229 .name = (char *) "server-internal",
230 .has_arg = required_argument,
234 { .name = (char *) "daemonize",
235 .has_arg = required_argument,
239 .name = (char *) "client",
240 .has_arg = required_argument,
244 .name = (char *) "remote-config",
245 .has_arg = required_argument,
249 .name = (char *) "cpuclock-test",
250 .has_arg = no_argument,
254 .name = (char *) "crctest",
255 .has_arg = optional_argument,
259 .name = (char *) "memcpytest",
260 .has_arg = optional_argument,
264 .name = (char *) "idle-prof",
265 .has_arg = required_argument,
269 .name = (char *) "status-interval",
270 .has_arg = required_argument,
271 .val = 'L' | FIO_CLIENT_FLAG,
274 .name = (char *) "trigger-file",
275 .has_arg = required_argument,
279 .name = (char *) "trigger-timeout",
280 .has_arg = required_argument,
284 .name = (char *) "trigger",
285 .has_arg = required_argument,
289 .name = (char *) "trigger-remote",
290 .has_arg = required_argument,
294 .name = (char *) "aux-path",
295 .has_arg = required_argument,
299 .name = (char *) "merge-blktrace-only",
300 .has_arg = no_argument,
301 .val = 'A' | FIO_CLIENT_FLAG,
308 void free_threads_shm(void)
312 for (i = 0; i < nr_segments; i++) {
313 struct thread_segment *seg = &segments[i];
316 void *tp = seg->threads;
317 #ifndef CONFIG_NO_SHM
318 struct shmid_ds sbuf;
322 shmctl(seg->shm_id, IPC_RMID, &sbuf);
335 static void free_shm(void)
337 #ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
340 fio_debug_jobp = NULL;
347 free(trigger_remote_cmd);
348 trigger_file = trigger_cmd = trigger_remote_cmd = NULL;
350 options_free(fio_options, &def_thread.o);
357 static int add_thread_segment(void)
359 struct thread_segment *seg = &segments[nr_segments];
360 size_t size = JOBS_PER_SEG * sizeof(struct thread_data);
363 if (nr_segments + 1 >= REAL_MAX_SEG) {
364 log_err("error: maximum number of jobs reached.\n");
368 size += 2 * sizeof(unsigned int);
370 #ifndef CONFIG_NO_SHM
371 seg->shm_id = shmget(0, size, IPC_CREAT | 0600);
372 if (seg->shm_id == -1) {
373 if (errno != EINVAL && errno != ENOMEM && errno != ENOSPC)
378 seg->threads = malloc(size);
383 #ifndef CONFIG_NO_SHM
384 seg->threads = shmat(seg->shm_id, NULL, 0);
385 if (seg->threads == (void *) -1) {
389 if (shm_attach_to_open_removed())
390 shmctl(seg->shm_id, IPC_RMID, NULL);
395 memset(seg->threads, 0, JOBS_PER_SEG * sizeof(struct thread_data));
396 for (i = 0; i < JOBS_PER_SEG; i++)
397 DRD_IGNORE_VAR(seg->threads[i]);
400 /* Not first segment, we're done */
401 if (nr_segments != 1) {
406 fio_debug_jobp = (unsigned int *)(seg->threads + JOBS_PER_SEG);
407 *fio_debug_jobp = -1;
408 fio_warned = fio_debug_jobp + 1;
416 * The thread areas are shared between the main process and the job
417 * threads/processes, and is split into chunks of JOBS_PER_SEG. If the current
418 * segment has no more room, add a new chunk.
420 static int expand_thread_area(void)
422 struct thread_segment *seg = &segments[cur_segment];
424 if (nr_segments && seg->nr_threads < JOBS_PER_SEG)
427 return add_thread_segment();
430 static void dump_print_option(struct print_option *p)
434 if (!strcmp("description", p->name))
439 log_info("--%s%s", p->name, p->value ? "" : " ");
441 log_info("=%s%s%s ", delim, p->value, delim);
444 static void dump_opt_list(struct thread_data *td)
446 struct flist_head *entry;
447 struct print_option *p;
449 if (flist_empty(&td->opt_list))
452 flist_for_each(entry, &td->opt_list) {
453 p = flist_entry(entry, struct print_option, list);
454 dump_print_option(p);
458 static void copy_opt_list(struct thread_data *dst, struct thread_data *src)
460 struct flist_head *entry;
462 if (flist_empty(&src->opt_list))
465 flist_for_each(entry, &src->opt_list) {
466 struct print_option *srcp, *dstp;
468 srcp = flist_entry(entry, struct print_option, list);
469 dstp = malloc(sizeof(*dstp));
470 dstp->name = strdup(srcp->name);
472 dstp->value = strdup(srcp->value);
475 flist_add_tail(&dstp->list, &dst->opt_list);
480 * Return a free job structure.
482 static struct thread_data *get_new_job(bool global, struct thread_data *parent,
483 bool preserve_eo, const char *jobname)
485 struct thread_segment *seg;
486 struct thread_data *td;
490 if (expand_thread_area()) {
491 log_err("error: failed to setup shm segment\n");
495 seg = &segments[cur_segment];
496 td = &seg->threads[seg->nr_threads++];
500 INIT_FLIST_HEAD(&td->opt_list);
501 if (parent != &def_thread)
502 copy_opt_list(td, parent);
509 td->o.uid = td->o.gid = -1U;
511 dup_files(td, parent);
512 fio_options_mem_dupe(td);
514 profile_add_hooks(td);
516 td->thread_number = thread_number;
517 td->subjob_number = 0;
520 td->o.name = strdup(jobname);
522 if (!parent->o.group_reporting || parent == &def_thread)
528 static void put_job(struct thread_data *td)
530 if (td == &def_thread)
537 log_info("fio: %s\n", td->verror);
539 fio_options_free(td);
540 fio_dump_options_free(td);
547 memset(td, 0, sizeof(*td));
548 segments[cur_segment].nr_threads--;
552 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
554 unsigned long long bs = td->o.min_bs[ddir];
556 assert(ddir_rw(ddir));
558 if (td->o.rate[ddir])
559 td->rate_bps[ddir] = td->o.rate[ddir];
561 td->rate_bps[ddir] = (uint64_t) td->o.rate_iops[ddir] * bs;
563 if (!td->rate_bps[ddir]) {
564 log_err("rate lower than supported\n");
568 td->rate_next_io_time[ddir] = 0;
569 td->rate_io_issue_bytes[ddir] = 0;
570 td->last_usec[ddir] = 0;
574 static int setup_rate(struct thread_data *td)
578 for_each_rw_ddir(ddir) {
579 if (td->o.rate[ddir] || td->o.rate_iops[ddir]) {
580 ret |= __setup_rate(td, ddir);
586 static int fixed_block_size(struct thread_options *o)
588 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
589 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
590 o->min_bs[DDIR_TRIM] == o->max_bs[DDIR_TRIM] &&
591 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE] &&
592 o->min_bs[DDIR_READ] == o->min_bs[DDIR_TRIM];
598 static unsigned int gcd(unsigned int m, unsigned int n)
603 return gcd(n, m % n);
607 * Lazy way of fixing up options that depend on each other. We could also
608 * define option callback handlers, but this is easier.
610 static int fixup_options(struct thread_data *td)
612 struct thread_options *o = &td->o;
615 if (read_only && (td_write(td) || td_trim(td))) {
616 log_err("fio: trim and write operations are not allowed"
617 " with the --readonly parameter.\n");
621 if (td_trimwrite(td) && o->num_range > 1) {
622 log_err("fio: trimwrite cannot be used with multiple"
627 if (td_trim(td) && o->num_range > 1 &&
628 !td_ioengine_flagged(td, FIO_MULTI_RANGE_TRIM)) {
629 log_err("fio: can't use multiple ranges with IO engine %s\n",
634 #ifndef CONFIG_PSHARED
635 if (!o->use_thread) {
636 log_info("fio: this platform does not support process shared"
637 " mutexes, forcing use of threads. Use the 'thread'"
638 " option to get rid of this warning.\n");
640 ret |= warnings_fatal;
644 if (o->write_iolog_file && o->read_iolog_file) {
645 log_err("fio: read iolog overrides write_iolog\n");
646 free(o->write_iolog_file);
647 o->write_iolog_file = NULL;
648 ret |= warnings_fatal;
651 if (o->zone_mode == ZONE_MODE_NONE && o->zone_size) {
652 log_err("fio: --zonemode=none and --zonesize are not compatible.\n");
656 if (o->zone_mode == ZONE_MODE_ZBD && !o->create_serialize) {
657 log_err("fio: --zonemode=zbd and --create_serialize=0 are not compatible.\n");
661 if (o->zone_mode == ZONE_MODE_STRIDED && !o->zone_size) {
662 log_err("fio: --zonesize must be specified when using --zonemode=strided.\n");
666 if (o->zone_mode == ZONE_MODE_NOT_SPECIFIED) {
668 o->zone_mode = ZONE_MODE_STRIDED;
670 o->zone_mode = ZONE_MODE_NONE;
674 * Strided zone mode only really works with 1 file.
676 if (o->zone_mode == ZONE_MODE_STRIDED && o->open_files > 1)
677 o->zone_mode = ZONE_MODE_NONE;
680 * If zone_range isn't specified, backward compatibility dictates it
681 * should be made equal to zone_size.
683 if (o->zone_mode == ZONE_MODE_STRIDED && !o->zone_range)
684 o->zone_range = o->zone_size;
687 * Reads can do overwrites, we always need to pre-create the file
692 for_each_rw_ddir(ddir) {
693 if (!o->min_bs[ddir])
694 o->min_bs[ddir] = o->bs[ddir];
695 if (!o->max_bs[ddir])
696 o->max_bs[ddir] = o->bs[ddir];
700 for_each_rw_ddir(ddir) {
701 o->rw_min_bs = min(o->rw_min_bs, o->min_bs[ddir]);
705 * For random IO, allow blockalign offset other than min_bs.
707 for_each_rw_ddir(ddir) {
708 if (!o->ba[ddir] || !td_random(td))
709 o->ba[ddir] = o->min_bs[ddir];
712 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
713 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE] ||
714 o->ba[DDIR_TRIM] != o->min_bs[DDIR_TRIM]) &&
716 log_err("fio: Any use of blockalign= turns off randommap\n");
718 ret |= warnings_fatal;
721 if (!o->file_size_high)
722 o->file_size_high = o->file_size_low;
724 if (o->start_delay_high) {
725 if (!o->start_delay_orig)
726 o->start_delay_orig = o->start_delay;
727 o->start_delay = rand_between(&td->delay_state,
729 o->start_delay_high);
732 if (o->norandommap && o->verify != VERIFY_NONE
733 && !fixed_block_size(o)) {
734 log_err("fio: norandommap given for variable block sizes, "
736 ret |= warnings_fatal;
738 if (o->bs_unaligned && (o->odirect || td_ioengine_flagged(td, FIO_RAWIO)))
739 log_err("fio: bs_unaligned may not work with raw io\n");
742 * thinktime_spin must be less than thinktime
744 if (o->thinktime_spin > o->thinktime)
745 o->thinktime_spin = o->thinktime;
748 * The low water mark cannot be bigger than the iodepth
750 if (o->iodepth_low > o->iodepth || !o->iodepth_low)
751 o->iodepth_low = o->iodepth;
754 * If batch number isn't set, default to the same as iodepth
756 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
757 o->iodepth_batch = o->iodepth;
760 * If max batch complete number isn't set or set incorrectly,
761 * default to the same as iodepth_batch_complete_min
763 if (o->iodepth_batch_complete_min > o->iodepth_batch_complete_max)
764 o->iodepth_batch_complete_max = o->iodepth_batch_complete_min;
767 * There's no need to check for in-flight overlapping IOs if the job
768 * isn't changing data or the maximum iodepth is guaranteed to be 1
769 * when we are not in offload mode
771 if (o->serialize_overlap && !(td->flags & TD_F_READ_IOLOG) &&
772 (!(td_write(td) || td_trim(td)) || o->iodepth == 1) &&
773 o->io_submit_mode != IO_MODE_OFFLOAD)
774 o->serialize_overlap = 0;
776 if (o->nr_files > td->files_index)
777 o->nr_files = td->files_index;
779 if (o->open_files > o->nr_files || !o->open_files)
780 o->open_files = o->nr_files;
782 if (((o->rate[DDIR_READ] + o->rate[DDIR_WRITE] + o->rate[DDIR_TRIM]) &&
783 (o->rate_iops[DDIR_READ] + o->rate_iops[DDIR_WRITE] + o->rate_iops[DDIR_TRIM])) ||
784 ((o->ratemin[DDIR_READ] + o->ratemin[DDIR_WRITE] + o->ratemin[DDIR_TRIM]) &&
785 (o->rate_iops_min[DDIR_READ] + o->rate_iops_min[DDIR_WRITE] + o->rate_iops_min[DDIR_TRIM]))) {
786 log_err("fio: rate and rate_iops are mutually exclusive\n");
789 for_each_rw_ddir(ddir) {
790 if ((o->rate[ddir] && (o->rate[ddir] < o->ratemin[ddir])) ||
791 (o->rate_iops[ddir] && (o->rate_iops[ddir] < o->rate_iops_min[ddir]))) {
792 log_err("fio: minimum rate exceeds rate, ddir %d\n", +ddir);
797 if (!o->timeout && o->time_based) {
798 log_err("fio: time_based requires a runtime/timeout setting\n");
800 ret |= warnings_fatal;
803 if (o->fill_device && !o->size)
806 if (o->verify != VERIFY_NONE) {
807 if (td_write(td) && o->do_verify && o->numjobs > 1 &&
809 !(o->unique_filename &&
810 strstr(o->filename_format, "$jobname") &&
811 strstr(o->filename_format, "$jobnum") &&
812 strstr(o->filename_format, "$filenum")))) {
813 log_info("fio: multiple writers may overwrite blocks "
814 "that belong to other jobs. This can cause "
815 "verification failures.\n");
816 ret |= warnings_fatal;
820 * Warn if verification is requested but no verification of any
821 * kind can be started due to time constraints
823 if (td_write(td) && o->do_verify && o->timeout &&
824 o->time_based && !td_read(td) && !o->verify_backlog) {
825 log_info("fio: verification read phase will never "
826 "start because write phase uses all of "
828 ret |= warnings_fatal;
831 if (!fio_option_is_set(o, refill_buffers))
832 o->refill_buffers = 1;
834 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
836 o->verify_interval = o->min_bs[DDIR_WRITE];
839 * Verify interval must be smaller or equal to the
842 if (o->verify_interval > o->min_bs[DDIR_WRITE])
843 o->verify_interval = o->min_bs[DDIR_WRITE];
844 else if (td_read(td) && o->verify_interval > o->min_bs[DDIR_READ])
845 o->verify_interval = o->min_bs[DDIR_READ];
848 * Verify interval must be a factor of both min and max
851 if (!o->verify_interval ||
852 (o->min_bs[DDIR_WRITE] % o->verify_interval) ||
853 (o->max_bs[DDIR_WRITE] % o->verify_interval))
854 o->verify_interval = gcd(o->min_bs[DDIR_WRITE],
855 o->max_bs[DDIR_WRITE]);
859 if (o->invalidate_cache)
860 o->invalidate_cache = 0;
861 if (td_ioengine_flagged(td, FIO_PIPEIO)) {
862 log_info("fio: cannot pre-read files with an IO engine"
863 " that isn't seekable. Pre-read disabled.\n");
864 ret |= warnings_fatal;
868 if (o->unit_base == N2S_NONE) {
869 if (td_ioengine_flagged(td, FIO_BIT_BASED))
870 o->unit_base = N2S_BITPERSEC;
872 o->unit_base = N2S_BYTEPERSEC;
875 #ifndef CONFIG_FDATASYNC
876 if (o->fdatasync_blocks) {
877 log_info("fio: this platform does not support fdatasync()"
878 " falling back to using fsync(). Use the 'fsync'"
879 " option instead of 'fdatasync' to get rid of"
881 o->fsync_blocks = o->fdatasync_blocks;
882 o->fdatasync_blocks = 0;
883 ret |= warnings_fatal;
889 * Windows doesn't support O_DIRECT or O_SYNC with the _open interface,
890 * so fail if we're passed those flags
892 if (td_ioengine_flagged(td, FIO_SYNCIO) && (o->odirect || o->sync_io)) {
893 log_err("fio: Windows does not support direct or non-buffered io with"
894 " the synchronous ioengines. Use the 'windowsaio' ioengine"
895 " with 'direct=1' and 'iodepth=1' instead.\n");
901 * For fully compressible data, just zero them at init time.
902 * It's faster than repeatedly filling it. For non-zero
903 * compression, we should have refill_buffers set. Set it, unless
904 * the job file already changed it.
906 if (o->compress_percentage) {
907 if (o->compress_percentage == 100) {
909 o->compress_percentage = 0;
910 } else if (!fio_option_is_set(o, refill_buffers)) {
911 o->refill_buffers = 1;
912 td->flags |= TD_F_REFILL_BUFFERS;
917 * Using a non-uniform random distribution excludes usage of
920 if (o->random_distribution != FIO_RAND_DIST_RANDOM)
924 * If size is set but less than the min block size, complain
926 if (o->size && o->size < td_min_bs(td)) {
927 log_err("fio: size too small, must not be less than minimum block size: %llu < %llu\n",
928 (unsigned long long) o->size, td_min_bs(td));
933 * If randseed is set, that overrides randrepeat
935 if (fio_option_is_set(o, rand_seed))
936 o->rand_repeatable = 0;
938 if (td_ioengine_flagged(td, FIO_NOEXTEND) && o->file_append) {
939 log_err("fio: can't append/extent with IO engine %s\n", td->io_ops->name);
943 if (fio_option_is_set(o, gtod_cpu)) {
945 fio_gtod_set_cpu(o->gtod_cpu);
946 fio_gtod_offload = 1;
949 td->loops = o->loops;
953 if (o->block_error_hist && o->nr_files != 1) {
954 log_err("fio: block error histogram only available "
955 "with a single file per job, but %d files "
956 "provided\n", o->nr_files);
961 o->lat_percentiles = 0;
963 o->clat_percentiles = 0;
965 o->slat_percentiles = 0;
967 /* Do this only for the parent job */
968 if (!td->subjob_number) {
970 * Fix these up to be nsec internally
972 for_each_rw_ddir(ddir)
973 o->max_latency[ddir] *= 1000ULL;
975 o->latency_target *= 1000ULL;
979 * Dedupe working set verifications
981 if (o->dedupe_percentage && o->dedupe_mode == DEDUPE_MODE_WORKING_SET) {
982 if (!fio_option_is_set(o, size)) {
983 log_err("fio: pregenerated dedupe working set "
984 "requires size to be set\n");
986 } else if (o->nr_files != 1) {
987 log_err("fio: dedupe working set mode supported with "
988 "single file per job, but %d files "
989 "provided\n", o->nr_files);
991 } else if (o->dedupe_working_set_percentage + o->dedupe_percentage > 100) {
992 log_err("fio: impossible to reach expected dedupe percentage %u "
993 "since %u percentage of size is reserved to dedupe working set "
994 "(those are unique pages)\n",
995 o->dedupe_percentage, o->dedupe_working_set_percentage);
1001 if (td->o.ss_check_interval != td2->o.ss_check_interval) {
1002 log_err("fio: conflicting ss_check_interval: %llu and %llu, must be globally equal\n",
1003 td->o.ss_check_interval, td2->o.ss_check_interval);
1007 if (td->o.ss_dur && td->o.ss_check_interval / 1000L < 1000) {
1008 log_err("fio: ss_check_interval must be at least 1s\n");
1012 if (td->o.ss_dur && (td->o.ss_dur % td->o.ss_check_interval != 0 || td->o.ss_dur <= td->o.ss_check_interval)) {
1013 log_err("fio: ss_duration %lluus must be multiple of ss_check_interval %lluus\n",
1014 td->o.ss_dur, td->o.ss_check_interval);
1019 if (fio_option_is_set(&td->o, dp_type) &&
1020 (td->o.dp_type == FIO_DP_STREAMS || td->o.dp_type == FIO_DP_NONE)) {
1021 log_err("fio: fdp=1 is not compatible with dataplacement={streams, none}\n");
1024 td->o.dp_type = FIO_DP_FDP;
1030 static void init_rand_file_service(struct thread_data *td)
1032 unsigned long nranges = td->o.nr_files << FIO_FSERVICE_SHIFT;
1033 const unsigned int seed = td->rand_seeds[FIO_RAND_FILE_OFF];
1035 if (td->o.file_service_type == FIO_FSERVICE_ZIPF) {
1036 zipf_init(&td->next_file_zipf, nranges, td->zipf_theta, td->random_center, seed);
1037 zipf_disable_hash(&td->next_file_zipf);
1038 } else if (td->o.file_service_type == FIO_FSERVICE_PARETO) {
1039 pareto_init(&td->next_file_zipf, nranges, td->pareto_h, td->random_center, seed);
1040 zipf_disable_hash(&td->next_file_zipf);
1041 } else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) {
1042 gauss_init(&td->next_file_gauss, nranges, td->gauss_dev, td->random_center, seed);
1043 gauss_disable_hash(&td->next_file_gauss);
1047 void td_fill_rand_seeds(struct thread_data *td)
1049 uint64_t read_seed = td->rand_seeds[FIO_RAND_BS_OFF];
1050 uint64_t write_seed = td->rand_seeds[FIO_RAND_BS1_OFF];
1051 uint64_t trim_seed = td->rand_seeds[FIO_RAND_BS2_OFF];
1055 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
1061 * trimwrite is special in that we need to generate the same
1062 * offsets to get the "write after trim" effect. If we are
1063 * using bssplit to set buffer length distributions, ensure that
1064 * we seed the trim and write generators identically. Ditto for
1065 * verify, read and writes must have the same seed, if we are doing
1068 if (td->o.verify != VERIFY_NONE)
1069 write_seed = read_seed;
1070 if (td_trimwrite(td))
1071 trim_seed = write_seed;
1072 init_rand_seed(&td->bsrange_state[DDIR_READ], read_seed, use64);
1073 init_rand_seed(&td->bsrange_state[DDIR_WRITE], write_seed, use64);
1074 init_rand_seed(&td->bsrange_state[DDIR_TRIM], trim_seed, use64);
1076 init_rand_seed(&td->verify_state, td->rand_seeds[FIO_RAND_VER_OFF],
1078 init_rand_seed(&td->rwmix_state, td->rand_seeds[FIO_RAND_MIX_OFF], false);
1080 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
1081 init_rand_seed(&td->next_file_state, td->rand_seeds[FIO_RAND_FILE_OFF], use64);
1082 else if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1083 init_rand_file_service(td);
1085 init_rand_seed(&td->file_size_state, td->rand_seeds[FIO_RAND_FILE_SIZE_OFF], use64);
1086 init_rand_seed(&td->trim_state, td->rand_seeds[FIO_RAND_TRIM_OFF], use64);
1087 init_rand_seed(&td->delay_state, td->rand_seeds[FIO_RAND_START_DELAY], use64);
1088 init_rand_seed(&td->poisson_state[0], td->rand_seeds[FIO_RAND_POISSON_OFF], 0);
1089 init_rand_seed(&td->poisson_state[1], td->rand_seeds[FIO_RAND_POISSON2_OFF], 0);
1090 init_rand_seed(&td->poisson_state[2], td->rand_seeds[FIO_RAND_POISSON3_OFF], 0);
1091 init_rand_seed(&td->dedupe_state, td->rand_seeds[FIO_DEDUPE_OFF], false);
1092 init_rand_seed(&td->zone_state, td->rand_seeds[FIO_RAND_ZONE_OFF], false);
1093 init_rand_seed(&td->prio_state, td->rand_seeds[FIO_RAND_PRIO_CMDS], false);
1094 init_rand_seed(&td->dedupe_working_set_index_state, td->rand_seeds[FIO_RAND_DEDUPE_WORKING_SET_IX], use64);
1096 init_rand_seed(&td->random_state, td->rand_seeds[FIO_RAND_BLOCK_OFF], use64);
1098 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1099 struct frand_state *s = &td->seq_rand_state[i];
1101 init_rand_seed(s, td->rand_seeds[FIO_RAND_SEQ_RAND_READ_OFF], false);
1104 init_rand_seed(&td->buf_state, td->rand_seeds[FIO_RAND_BUF_OFF], use64);
1105 frand_copy(&td->buf_state_prev, &td->buf_state);
1107 init_rand_seed(&td->fdp_state, td->rand_seeds[FIO_RAND_FDP_OFF], use64);
1110 static int setup_random_seeds(struct thread_data *td)
1115 if (!td->o.rand_repeatable && !fio_option_is_set(&td->o, rand_seed)) {
1116 int ret = init_random_seeds(td->rand_seeds, sizeof(td->rand_seeds));
1117 dprint(FD_RANDOM, "using system RNG for random seeds\n");
1121 seed = td->o.rand_seed;
1122 for (i = 0; i < 4; i++)
1123 seed *= 0x9e370001UL;
1125 for (i = 0; i < FIO_RAND_NR_OFFS; i++) {
1126 td->rand_seeds[i] = seed * td->thread_number + i;
1127 seed *= 0x9e370001UL;
1131 td_fill_rand_seeds(td);
1133 dprint(FD_RANDOM, "FIO_RAND_NR_OFFS=%d\n", FIO_RAND_NR_OFFS);
1134 for (int i = 0; i < FIO_RAND_NR_OFFS; i++)
1135 dprint(FD_RANDOM, "rand_seeds[%d]=%" PRIu64 "\n", i, td->rand_seeds[i]);
1141 * Initializes the ioengine configured for a job, if it has not been done so
1144 int ioengine_load(struct thread_data *td)
1146 if (!td->o.ioengine) {
1147 log_err("fio: internal fault, no IO engine specified\n");
1152 struct ioengine_ops *ops;
1155 /* An engine is loaded, but the requested ioengine
1158 if (!strcmp(td->io_ops->name, td->o.ioengine)) {
1159 /* The right engine is already loaded */
1164 * Name of file and engine may be different, load ops
1165 * for this name and see if they match. If they do, then
1166 * the engine is unchanged.
1168 dlhandle = td->io_ops->dlhandle;
1169 ops = load_ioengine(td);
1173 if (ops == td->io_ops && dlhandle == td->io_ops->dlhandle)
1176 if (dlhandle && dlhandle != td->io_ops->dlhandle)
1179 /* Unload the old engine. */
1183 td->io_ops = load_ioengine(td);
1187 if (td->io_ops->option_struct_size && td->io_ops->options) {
1189 * In cases where td->eo is set, clone it for a child thread.
1190 * This requires that the parent thread has the same ioengine,
1191 * but that requirement must be enforced by the code which
1192 * cloned the thread.
1194 void *origeo = td->eo;
1196 * Otherwise use the default thread options.
1198 if (!origeo && td != &def_thread && def_thread.eo &&
1199 def_thread.io_ops->options == td->io_ops->options)
1200 origeo = def_thread.eo;
1202 options_init(td->io_ops->options);
1203 td->eo = malloc(td->io_ops->option_struct_size);
1205 * Use the default thread as an option template if this uses the
1206 * same options structure and there are non-default options
1210 memcpy(td->eo, origeo, td->io_ops->option_struct_size);
1211 options_mem_dupe(td->io_ops->options, td->eo);
1213 memset(td->eo, 0, td->io_ops->option_struct_size);
1214 fill_default_options(td->eo, td->io_ops->options);
1216 *(struct thread_data **)td->eo = td;
1220 td->io_ops->flags |= FIO_RAWIO;
1222 td_set_ioengine_flags(td);
1226 log_err("fio: failed to load engine\n");
1231 static void init_flags(struct thread_data *td)
1233 struct thread_options *o = &td->o;
1236 if (o->verify_backlog)
1237 td->flags |= TD_F_VER_BACKLOG;
1238 if (o->trim_backlog)
1239 td->flags |= TD_F_TRIM_BACKLOG;
1240 if (o->read_iolog_file)
1241 td->flags |= TD_F_READ_IOLOG;
1242 if (o->refill_buffers)
1243 td->flags |= TD_F_REFILL_BUFFERS;
1245 * Always scramble buffers if asked to
1247 if (o->scramble_buffers && fio_option_is_set(o, scramble_buffers))
1248 td->flags |= TD_F_SCRAMBLE_BUFFERS;
1250 * But also scramble buffers, unless we were explicitly asked
1253 if (o->scramble_buffers && !(o->zero_buffers &&
1254 fio_option_is_set(o, zero_buffers)))
1255 td->flags |= TD_F_SCRAMBLE_BUFFERS;
1256 if (o->verify != VERIFY_NONE)
1257 td->flags |= TD_F_DO_VERIFY;
1259 if (o->verify_async || o->io_submit_mode == IO_MODE_OFFLOAD)
1260 td->flags |= TD_F_NEED_LOCK;
1262 if (o->mem_type == MEM_CUDA_MALLOC)
1263 td->flags &= ~TD_F_SCRAMBLE_BUFFERS;
1265 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1266 if (option_check_rate(td, i)) {
1267 td->flags |= TD_F_CHECK_RATE;
1281 static struct fpre_keyword {
1282 const char *keyword;
1285 } fpre_keywords[] = {
1286 { .keyword = "$jobname", .key = FPRE_JOBNAME, },
1287 { .keyword = "$jobnum", .key = FPRE_JOBNUM, },
1288 { .keyword = "$filenum", .key = FPRE_FILENUM, },
1289 { .keyword = "$clientuid", .key = FPRE_CLIENTUID, },
1290 { .keyword = NULL, },
1293 static char *make_filename(char *buf, size_t buf_size,struct thread_options *o,
1294 const char *jobname, int jobnum, int filenum)
1296 struct fpre_keyword *f;
1297 char copy[PATH_MAX];
1298 size_t dst_left = PATH_MAX - 1;
1300 if (!o->filename_format || !strlen(o->filename_format)) {
1301 sprintf(buf, "%s.%d.%d", jobname, jobnum, filenum);
1305 for (f = &fpre_keywords[0]; f->keyword; f++)
1306 f->strlen = strlen(f->keyword);
1308 snprintf(buf, buf_size, "%s", o->filename_format);
1310 memset(copy, 0, sizeof(copy));
1311 for (f = &fpre_keywords[0]; f->keyword; f++) {
1313 size_t pre_len, post_start = 0;
1314 char *str, *dst = copy;
1316 str = strcasestr(buf, f->keyword);
1320 pre_len = str - buf;
1321 if (strlen(str) != f->strlen)
1322 post_start = pre_len + f->strlen;
1325 strncpy(dst, buf, pre_len);
1327 dst_left -= pre_len;
1331 case FPRE_JOBNAME: {
1334 ret = snprintf(dst, dst_left, "%s", jobname);
1337 else if (ret > dst_left) {
1338 log_err("fio: truncated filename\n");
1350 ret = snprintf(dst, dst_left, "%d", jobnum);
1353 else if (ret > dst_left) {
1354 log_err("fio: truncated filename\n");
1363 case FPRE_FILENUM: {
1366 ret = snprintf(dst, dst_left, "%d", filenum);
1369 else if (ret > dst_left) {
1370 log_err("fio: truncated filename\n");
1379 case FPRE_CLIENTUID: {
1381 ret = snprintf(dst, dst_left, "%s", client_sockaddr_str);
1384 else if (ret > dst_left) {
1385 log_err("fio: truncated filename\n");
1400 strncpy(dst, buf + post_start, dst_left);
1402 snprintf(buf, buf_size, "%s", copy);
1409 bool parse_dryrun(void)
1411 return dump_cmdline || parse_only;
1414 static void gen_log_name(char *name, size_t size, const char *logtype,
1415 const char *logname, unsigned int num,
1416 const char *suf, int per_job)
1419 snprintf(name, size, "%s_%s.%d.%s", logname, logtype, num, suf);
1421 snprintf(name, size, "%s_%s.%s", logname, logtype, suf);
1424 static int check_waitees(char *waitee)
1429 if (td->subjob_number)
1432 ret += !strcmp(td->o.name, waitee);
1438 static bool wait_for_ok(const char *jobname, struct thread_options *o)
1445 if (!strcmp(jobname, o->wait_for)) {
1446 log_err("%s: a job cannot wait for itself (wait_for=%s).\n",
1447 jobname, o->wait_for);
1451 if (!(nw = check_waitees(o->wait_for))) {
1452 log_err("%s: waitee job %s unknown.\n", jobname, o->wait_for);
1457 log_err("%s: multiple waitees %s found,\n"
1458 "please avoid duplicates when using wait_for option.\n",
1459 jobname, o->wait_for);
1466 static int verify_per_group_options(struct thread_data *td, const char *jobname)
1469 if (td->groupid != td2->groupid)
1473 td->o.lat_percentiles != td2->o.lat_percentiles) {
1474 log_err("fio: lat_percentiles in job: %s differs from group\n",
1484 * Treat an empty log file name the same as a one not given
1486 static const char *make_log_name(const char *logname, const char *jobname)
1488 if (logname && strcmp(logname, ""))
1495 * Adds a job to the list of things todo. Sanitizes the various options
1496 * to make sure we don't have conflicts, and initializes various
1499 static int add_job(struct thread_data *td, const char *jobname, int job_add_num,
1500 int recursed, int client_type)
1503 char fname[PATH_MAX + 1];
1504 int numjobs, file_alloced;
1505 struct thread_options *o = &td->o;
1506 char logname[PATH_MAX + 32];
1509 * the def_thread is just for options, it's not a real job
1511 if (td == &def_thread)
1517 * if we are just dumping the output command line, don't add the job
1519 if (parse_dryrun()) {
1524 td->client_type = client_type;
1526 if (profile_td_init(td))
1529 if (ioengine_load(td))
1533 if (!o->filename && !td->files_index && !o->read_iolog_file) {
1536 if (o->nr_files == 1 && exists_and_not_regfile(jobname))
1537 add_file(td, jobname, job_add_num, 0);
1539 for (i = 0; i < o->nr_files; i++)
1540 add_file(td, make_filename(fname, sizeof(fname), o, jobname, job_add_num, i), job_add_num, 0);
1544 if (setup_random_seeds(td)) {
1545 td_verror(td, errno, "setup_random_seeds");
1549 if (fixup_options(td))
1552 if (!td->o.dedupe_global && init_dedupe_working_set_seeds(td, 0))
1556 * Belongs to fixup_options, but o->name is not necessarily set as yet
1558 if (!wait_for_ok(jobname, o))
1564 * IO engines only need this for option callbacks, and the address may
1565 * change in subprocesses.
1568 *(struct thread_data **)td->eo = NULL;
1570 if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1573 for_each_file(td, f, i)
1574 f->real_file_size = -1ULL;
1577 td->sem = fio_sem_init(FIO_SEM_LOCKED);
1579 td->ts.clat_percentiles = o->clat_percentiles;
1580 td->ts.lat_percentiles = o->lat_percentiles;
1581 td->ts.slat_percentiles = o->slat_percentiles;
1582 td->ts.percentile_precision = o->percentile_precision;
1583 memcpy(td->ts.percentile_list, o->percentile_list, sizeof(o->percentile_list));
1584 td->ts.sig_figs = o->sig_figs;
1586 init_thread_stat_min_vals(&td->ts);
1589 * td->>ddir_seq_nr needs to be initialized to 1, NOT o->ddir_seq_nr,
1590 * so that get_next_offset gets a new random offset the first time it
1591 * is called, instead of keeping an initial offset of 0 for the first
1594 td->ddir_seq_nr = 1;
1596 if ((o->stonewall || o->new_group) && prev_group_jobs) {
1597 prev_group_jobs = 0;
1599 if (groupid == INT_MAX) {
1600 log_err("fio: too many groups defined\n");
1605 td->groupid = groupid;
1608 if (td->o.group_reporting && prev_group_jobs > 1 &&
1609 verify_per_group_options(td, jobname))
1615 if (o->write_lat_log) {
1616 struct log_params p = {
1618 .avg_msec = o->log_avg_msec,
1619 .hist_msec = o->log_hist_msec,
1620 .hist_coarseness = o->log_hist_coarseness,
1621 .log_type = IO_LOG_TYPE_LAT,
1622 .log_offset = o->log_offset,
1623 .log_prio = o->log_prio,
1624 .log_gz = o->log_gz,
1625 .log_gz_store = o->log_gz_store,
1627 const char *pre = make_log_name(o->lat_log_file, o->name);
1635 if (!o->disable_lat) {
1636 gen_log_name(logname, sizeof(logname), "lat", pre,
1637 td->thread_number, suf, o->per_job_logs);
1638 setup_log(&td->lat_log, &p, logname);
1641 if (!o->disable_slat) {
1642 gen_log_name(logname, sizeof(logname), "slat", pre,
1643 td->thread_number, suf, o->per_job_logs);
1644 setup_log(&td->slat_log, &p, logname);
1647 if (!o->disable_clat) {
1648 gen_log_name(logname, sizeof(logname), "clat", pre,
1649 td->thread_number, suf, o->per_job_logs);
1650 setup_log(&td->clat_log, &p, logname);
1655 if (o->write_hist_log) {
1656 struct log_params p = {
1658 .avg_msec = o->log_avg_msec,
1659 .hist_msec = o->log_hist_msec,
1660 .hist_coarseness = o->log_hist_coarseness,
1661 .log_type = IO_LOG_TYPE_HIST,
1662 .log_offset = o->log_offset,
1663 .log_prio = o->log_prio,
1664 .log_gz = o->log_gz,
1665 .log_gz_store = o->log_gz_store,
1667 const char *pre = make_log_name(o->hist_log_file, o->name);
1671 if (td->client_type) {
1672 log_err("fio: --write_hist_log requires zlib in client/server mode\n");
1682 gen_log_name(logname, sizeof(logname), "clat_hist", pre,
1683 td->thread_number, suf, o->per_job_logs);
1684 setup_log(&td->clat_hist_log, &p, logname);
1687 if (o->write_bw_log) {
1688 struct log_params p = {
1690 .avg_msec = o->log_avg_msec,
1691 .hist_msec = o->log_hist_msec,
1692 .hist_coarseness = o->log_hist_coarseness,
1693 .log_type = IO_LOG_TYPE_BW,
1694 .log_offset = o->log_offset,
1695 .log_prio = o->log_prio,
1696 .log_gz = o->log_gz,
1697 .log_gz_store = o->log_gz_store,
1699 const char *pre = make_log_name(o->bw_log_file, o->name);
1702 if (fio_option_is_set(o, bw_avg_time))
1703 p.avg_msec = min(o->log_avg_msec, o->bw_avg_time);
1705 o->bw_avg_time = p.avg_msec;
1707 p.hist_msec = o->log_hist_msec;
1708 p.hist_coarseness = o->log_hist_coarseness;
1715 gen_log_name(logname, sizeof(logname), "bw", pre,
1716 td->thread_number, suf, o->per_job_logs);
1717 setup_log(&td->bw_log, &p, logname);
1719 if (o->write_iops_log) {
1720 struct log_params p = {
1722 .avg_msec = o->log_avg_msec,
1723 .hist_msec = o->log_hist_msec,
1724 .hist_coarseness = o->log_hist_coarseness,
1725 .log_type = IO_LOG_TYPE_IOPS,
1726 .log_offset = o->log_offset,
1727 .log_prio = o->log_prio,
1728 .log_gz = o->log_gz,
1729 .log_gz_store = o->log_gz_store,
1731 const char *pre = make_log_name(o->iops_log_file, o->name);
1734 if (fio_option_is_set(o, iops_avg_time))
1735 p.avg_msec = min(o->log_avg_msec, o->iops_avg_time);
1737 o->iops_avg_time = p.avg_msec;
1739 p.hist_msec = o->log_hist_msec;
1740 p.hist_coarseness = o->log_hist_coarseness;
1747 gen_log_name(logname, sizeof(logname), "iops", pre,
1748 td->thread_number, suf, o->per_job_logs);
1749 setup_log(&td->iops_log, &p, logname);
1753 o->name = strdup(jobname);
1755 if (output_format & FIO_OUTPUT_NORMAL) {
1757 if (is_backend && !recursed)
1758 fio_server_send_add_job(td);
1760 if (!td_ioengine_flagged(td, FIO_NOIO)) {
1761 char *c1, *c2, *c3, *c4;
1762 char *c5 = NULL, *c6 = NULL;
1763 int i2p = is_power_of_2(o->kb_base);
1764 struct buf_output out;
1766 c1 = num2str(o->min_bs[DDIR_READ], o->sig_figs, 1, i2p, N2S_BYTE);
1767 c2 = num2str(o->max_bs[DDIR_READ], o->sig_figs, 1, i2p, N2S_BYTE);
1768 c3 = num2str(o->min_bs[DDIR_WRITE], o->sig_figs, 1, i2p, N2S_BYTE);
1769 c4 = num2str(o->max_bs[DDIR_WRITE], o->sig_figs, 1, i2p, N2S_BYTE);
1771 if (!o->bs_is_seq_rand) {
1772 c5 = num2str(o->min_bs[DDIR_TRIM], o->sig_figs, 1, i2p, N2S_BYTE);
1773 c6 = num2str(o->max_bs[DDIR_TRIM], o->sig_figs, 1, i2p, N2S_BYTE);
1776 buf_output_init(&out);
1777 __log_buf(&out, "%s: (g=%d): rw=%s, ", td->o.name,
1779 ddir_str(o->td_ddir));
1781 if (o->bs_is_seq_rand)
1782 __log_buf(&out, "bs=(R) %s-%s, (W) %s-%s, bs_is_seq_rand, ",
1785 __log_buf(&out, "bs=(R) %s-%s, (W) %s-%s, (T) %s-%s, ",
1786 c1, c2, c3, c4, c5, c6);
1788 __log_buf(&out, "ioengine=%s, iodepth=%u\n",
1789 td->io_ops->name, o->iodepth);
1790 log_info_buf(out.buf, out.buflen);
1791 buf_output_free(&out);
1800 } else if (job_add_num == 1)
1804 if (td_steadystate_init(td))
1807 if (o->merge_blktrace_file && !merge_blktrace_iologs(td))
1810 if (merge_blktrace_only) {
1816 * recurse add identical jobs, clear numjobs and stonewall options
1817 * as they don't apply to sub-jobs
1819 numjobs = o->numjobs;
1821 struct thread_data *td_new = get_new_job(false, td, true, jobname);
1826 td_new->o.numjobs = 1;
1827 td_new->o.stonewall = 0;
1828 td_new->o.new_group = 0;
1829 td_new->subjob_number = numjobs;
1830 td_new->o.ss_dur = o->ss_dur * 1000000l;
1831 td_new->o.ss_limit = o->ss_limit;
1834 if (td_new->files) {
1836 for_each_file(td_new, f, i)
1838 free(td_new->files);
1839 td_new->files = NULL;
1841 td_new->files_index = 0;
1842 td_new->files_size = 0;
1843 if (td_new->o.filename) {
1844 free(td_new->o.filename);
1845 td_new->o.filename = NULL;
1849 if (add_job(td_new, jobname, numjobs, 1, client_type))
1860 * Parse as if 'o' was a command line
1862 void add_job_opts(const char **o, int client_type)
1864 struct thread_data *td, *td_parent;
1865 int i, in_global = 1;
1869 td_parent = td = NULL;
1871 if (!strncmp(o[i], "name", 4)) {
1874 add_job(td, jobname, 0, 0, client_type);
1876 sprintf(jobname, "%s", o[i] + 5);
1878 if (in_global && !td_parent)
1879 td_parent = get_new_job(true, &def_thread, false, jobname);
1880 else if (!in_global && !td) {
1882 td_parent = &def_thread;
1883 td = get_new_job(false, td_parent, false, jobname);
1886 fio_options_parse(td_parent, (char **) &o[i], 1);
1888 fio_options_parse(td, (char **) &o[i], 1);
1893 add_job(td, jobname, 0, 0, client_type);
1896 static int skip_this_section(const char *name)
1900 if (!nr_job_sections)
1902 if (!strncmp(name, "global", 6))
1905 for (i = 0; i < nr_job_sections; i++)
1906 if (!strcmp(job_sections[i], name))
1912 static int is_empty_or_comment(char *line)
1916 for (i = 0; i < strlen(line); i++) {
1921 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
1929 * This is our [ini] type file parser.
1931 static int __parse_jobs_ini(struct thread_data *td,
1932 char *file, int is_buf, int stonewall_flag, int type,
1933 int nested, char *name, char ***popts, int *aopts, int *nopts)
1935 bool global = false;
1936 bool stdin_occupied = false;
1940 int ret = 0, stonewall;
1943 int inside_skip = 0;
1945 int i, alloc_opts, num_opts;
1947 dprint(FD_PARSE, "Parsing ini file %s\n", file);
1948 assert(td || !nested);
1953 if (!strcmp(file, "-")) {
1955 stdin_occupied = true;
1957 f = fopen(file, "r");
1962 log_err("fio: unable to open '%s' job file\n", file);
1964 td_verror(td, __err, "job file open");
1969 string = malloc(OPT_LEN_MAX);
1972 * it's really 256 + small bit, 280 should suffice
1975 name = calloc(1, 280);
1979 if (nested && popts) {
1981 alloc_opts = *aopts;
1987 opts = malloc(sizeof(char *) * alloc_opts);
1991 stonewall = stonewall_flag;
1994 * if skip_fgets is set, we already have loaded a line we
1999 p = strsep(&file, "\n");
2001 p = fgets(string, OPT_LEN_MAX, f);
2007 strip_blank_front(&p);
2010 dprint(FD_PARSE, "%s\n", p);
2011 if (is_empty_or_comment(p))
2015 if (sscanf(p, "[%255[^\n]]", name) != 1) {
2019 log_err("fio: option <%s> outside of "
2020 "[] job section\n", p);
2025 name[strlen(name) - 1] = '\0';
2027 if (skip_this_section(name)) {
2033 dprint(FD_PARSE, "Parsing section [%s]\n", name);
2035 global = !strncmp(name, "global", 6);
2041 log_info("--name=%s ", name);
2045 td = get_new_job(global, &def_thread, false, name);
2052 * Separate multiple job files by a stonewall
2054 if (!global && stonewall) {
2055 td->o.stonewall = stonewall;
2060 memset(opts, 0, alloc_opts * sizeof(char *));
2068 p = strsep(&file, "\n");
2070 p = fgets(string, OPT_LEN_MAX, f);
2073 dprint(FD_PARSE, "%s", p);
2078 if (is_empty_or_comment(p))
2081 strip_blank_front(&p);
2084 * new section, break out and make sure we don't
2085 * fgets() a new line at the top.
2089 log_err("No new sections in included files\n");
2100 if (!strncmp(p, "include", strlen("include"))) {
2101 char *filename = p + strlen("include") + 1,
2102 *ts, *full_fn = NULL;
2105 * Allow for the include filename
2106 * specification to be relative.
2108 if (access(filename, F_OK) &&
2109 (ts = strrchr(file, '/'))) {
2110 if (asprintf(&full_fn, "%.*s%s",
2111 (int)(ts - file + 1), file,
2119 ret = __parse_jobs_ini(td, filename, is_buf,
2120 stonewall_flag, type, 1,
2122 &alloc_opts, &num_opts);
2125 log_err("Error %d while parsing "
2126 "include file %s\n",
2139 if (num_opts == alloc_opts) {
2141 opts = realloc(opts,
2142 alloc_opts * sizeof(char *));
2145 opts[num_opts] = strdup(p);
2151 *aopts = alloc_opts;
2156 ret = fio_options_parse(td, opts, num_opts);
2158 if (!ret && td->o.read_iolog_file != NULL) {
2159 char *fname = get_name_by_idx(td->o.read_iolog_file,
2161 if (!strcmp(fname, "-")) {
2162 if (stdin_occupied) {
2163 log_err("fio: only one user (read_iolog_file/job "
2164 "file) of stdin is permitted at once but "
2165 "more than one was found.\n");
2168 stdin_occupied = true;
2175 ret = add_job(td, name, 0, 0, type);
2177 log_err("fio: job %s dropped\n", name);
2181 for (i = 0; i < num_opts; i++)
2190 while (i < nr_job_sections) {
2191 free(job_sections[i]);
2196 job_sections = NULL;
2197 nr_job_sections = 0;
2204 if (!is_buf && f != stdin)
2209 int parse_jobs_ini(char *file, int is_buf, int stonewall_flag, int type)
2211 return __parse_jobs_ini(NULL, file, is_buf, stonewall_flag, type,
2212 0, NULL, NULL, NULL, NULL);
2215 static int fill_def_thread(void)
2217 memset(&def_thread, 0, sizeof(def_thread));
2218 INIT_FLIST_HEAD(&def_thread.opt_list);
2220 fio_getaffinity(getpid(), &def_thread.o.cpumask);
2221 def_thread.o.error_dump = 1;
2224 * fill default options
2226 fio_fill_default_options(&def_thread);
2230 static void show_debug_categories(void)
2232 #ifdef FIO_INC_DEBUG
2233 const struct debug_level *dl = &debug_levels[0];
2234 int curlen, first = 1;
2238 int has_next = (dl + 1)->name != NULL;
2240 if (first || curlen + strlen(dl->name) >= 80) {
2245 curlen += printf("\t\t\t%s", dl->name);
2246 curlen += 3 * (8 - 1);
2248 curlen += printf(",");
2250 curlen += printf("%s", dl->name);
2252 curlen += printf(",");
2262 * Following options aren't printed by usage().
2263 * --append-terse - Equivalent to --output-format=terse, see f6a7df53.
2264 * --latency-log - Deprecated option.
2266 static void usage(const char *name)
2268 printf("%s\n", fio_version_string);
2269 printf("%s [options] [job options] <job file(s)>\n", name);
2270 printf(" --debug=options\tEnable debug logging. May be one/more of:\n");
2271 show_debug_categories();
2272 printf(" --parse-only\t\tParse options only, don't start any IO\n");
2273 printf(" --merge-blktrace-only\tMerge blktraces only, don't start any IO\n");
2274 printf(" --output\t\tWrite output to file\n");
2275 printf(" --bandwidth-log\tGenerate aggregate bandwidth logs\n");
2276 printf(" --minimal\t\tMinimal (terse) output\n");
2277 printf(" --output-format=type\tOutput format (terse,json,json+,normal)\n");
2278 printf(" --terse-version=type\tSet terse version output format"
2279 " (default 3, or 2 or 4 or 5)\n");
2280 printf(" --version\t\tPrint version info and exit\n");
2281 printf(" --help\t\tPrint this page\n");
2282 printf(" --cpuclock-test\tPerform test/validation of CPU clock\n");
2283 printf(" --crctest=[type]\tTest speed of checksum functions\n");
2284 printf(" --cmdhelp=cmd\t\tPrint command help, \"all\" for all of"
2286 printf(" --enghelp=engine\tPrint ioengine help, or list"
2287 " available ioengines\n");
2288 printf(" --enghelp=engine,cmd\tPrint help for an ioengine"
2290 printf(" --showcmd\t\tTurn a job file into command line options\n");
2291 printf(" --eta=when\t\tWhen ETA estimate should be printed\n");
2292 printf(" \t\tMay be \"always\", \"never\" or \"auto\"\n");
2293 printf(" --eta-newline=t\tForce a new line for every 't'");
2294 printf(" period passed\n");
2295 printf(" --status-interval=t\tForce full status dump every");
2296 printf(" 't' period passed\n");
2297 printf(" --readonly\t\tTurn on safety read-only checks, preventing"
2299 printf(" --section=name\tOnly run specified section in job file,"
2300 " multiple sections can be specified\n");
2301 printf(" --alloc-size=kb\tSet smalloc pool to this size in kb"
2303 printf(" --warnings-fatal\tFio parser warnings are fatal\n");
2304 printf(" --max-jobs=nr\t\tMaximum number of threads/processes to support\n");
2305 printf(" --server=args\t\tStart a backend fio server\n");
2306 printf(" --daemonize=pidfile\tBackground fio server, write pid to file\n");
2307 printf(" --client=hostname\tTalk to remote backend(s) fio server at hostname\n");
2308 printf(" --remote-config=file\tTell fio server to load this local job file\n");
2309 printf(" --idle-prof=option\tReport cpu idleness on a system or percpu basis\n"
2310 "\t\t\t(option=system,percpu) or run unit work\n"
2311 "\t\t\tcalibration only (option=calibrate)\n");
2313 printf(" --inflate-log=log\tInflate and output compressed log\n");
2315 printf(" --trigger-file=file\tExecute trigger cmd when file exists\n");
2316 printf(" --trigger-timeout=t\tExecute trigger at this time\n");
2317 printf(" --trigger=cmd\t\tSet this command as local trigger\n");
2318 printf(" --trigger-remote=cmd\tSet this command as remote trigger\n");
2319 printf(" --aux-path=path\tUse this path for fio state generated files\n");
2320 printf("\nFio was written by Jens Axboe <axboe@kernel.dk>\n");
2323 #ifdef FIO_INC_DEBUG
2324 const struct debug_level debug_levels[] = {
2325 { .name = "process",
2326 .help = "Process creation/exit logging",
2327 .shift = FD_PROCESS,
2330 .help = "File related action logging",
2334 .help = "IO and IO engine action logging (offsets, queue, completions, etc)",
2338 .help = "Memory allocation/freeing logging",
2341 { .name = "blktrace",
2342 .help = "blktrace action logging",
2343 .shift = FD_BLKTRACE,
2346 .help = "IO verification action logging",
2350 .help = "Random generation logging",
2354 .help = "Parser logging",
2357 { .name = "diskutil",
2358 .help = "Disk utility logging actions",
2359 .shift = FD_DISKUTIL,
2362 .help = "Logging related to creating/destroying jobs",
2366 .help = "Mutex logging",
2369 { .name = "profile",
2370 .help = "Logging related to profiles",
2371 .shift = FD_PROFILE,
2374 .help = "Logging related to time keeping functions",
2378 .help = "Network logging",
2382 .help = "Rate logging",
2385 { .name = "compress",
2386 .help = "Log compression logging",
2387 .shift = FD_COMPRESS,
2389 { .name = "steadystate",
2390 .help = "Steady state detection logging",
2391 .shift = FD_STEADYSTATE,
2393 { .name = "helperthread",
2394 .help = "Helper thread logging",
2395 .shift = FD_HELPERTHREAD,
2398 .help = "Zoned Block Device logging",
2404 static int set_debug(const char *string)
2406 const struct debug_level *dl;
2407 char *p = (char *) string;
2414 if (!strcmp(string, "?") || !strcmp(string, "help")) {
2415 log_info("fio: dumping debug options:");
2416 for (i = 0; debug_levels[i].name; i++) {
2417 dl = &debug_levels[i];
2418 log_info("%s,", dl->name);
2424 while ((opt = strsep(&p, ",")) != NULL) {
2427 if (!strncmp(opt, "all", 3)) {
2428 log_info("fio: set all debug options\n");
2433 for (i = 0; debug_levels[i].name; i++) {
2434 dl = &debug_levels[i];
2435 found = !strncmp(opt, dl->name, strlen(dl->name));
2439 if (dl->shift == FD_JOB) {
2440 opt = strchr(opt, ':');
2442 log_err("fio: missing job number\n");
2446 fio_debug_jobno = atoi(opt);
2447 log_info("fio: set debug jobno %d\n",
2450 log_info("fio: set debug option %s\n", opt);
2451 fio_debug |= (1UL << dl->shift);
2457 log_err("fio: debug mask %s not found\n", opt);
2462 static int set_debug(const char *string)
2464 log_err("fio: debug tracing not included in build\n");
2469 static void fio_options_fill_optstring(void)
2471 char *ostr = cmd_optstr;
2475 while (l_opts[i].name) {
2476 ostr[c++] = l_opts[i].val;
2477 if (l_opts[i].has_arg == required_argument)
2479 else if (l_opts[i].has_arg == optional_argument) {
2488 static int client_flag_set(char c)
2493 while (l_opts[i].name) {
2494 int val = l_opts[i].val;
2496 if (c == (val & 0xff))
2497 return (val & FIO_CLIENT_FLAG);
2505 static void parse_cmd_client(void *client, char *opt)
2507 fio_client_add_cmd_option(client, opt);
2510 static void show_closest_option(const char *name)
2512 int best_option, best_distance;
2515 while (*name == '-')
2519 best_distance = INT_MAX;
2521 while (l_opts[i].name) {
2522 distance = string_distance(name, l_opts[i].name);
2523 if (distance < best_distance) {
2524 best_distance = distance;
2530 if (best_option != -1 && string_distance_ok(name, best_distance))
2531 log_err("Did you mean %s?\n", l_opts[best_option].name);
2534 static int parse_output_format(const char *optarg)
2536 char *p, *orig, *opt;
2539 p = orig = strdup(optarg);
2543 while ((opt = strsep(&p, ",")) != NULL) {
2544 if (!strcmp(opt, "minimal") ||
2545 !strcmp(opt, "terse") ||
2546 !strcmp(opt, "csv"))
2547 output_format |= FIO_OUTPUT_TERSE;
2548 else if (!strcmp(opt, "json"))
2549 output_format |= FIO_OUTPUT_JSON;
2550 else if (!strcmp(opt, "json+"))
2551 output_format |= (FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS);
2552 else if (!strcmp(opt, "normal"))
2553 output_format |= FIO_OUTPUT_NORMAL;
2555 log_err("fio: invalid output format %s\n", opt);
2565 int parse_cmd_line(int argc, char *argv[], int client_type)
2567 struct thread_data *td = NULL;
2568 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
2569 char *ostr = cmd_optstr;
2570 char *pid_file = NULL;
2571 void *cur_client = NULL;
2572 bool backend = false;
2575 * Reset optind handling, since we may call this multiple times
2580 while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
2581 if ((c & FIO_CLIENT_FLAG) || client_flag_set(c)) {
2582 parse_cmd_client(cur_client, argv[optind - 1]);
2583 c &= ~FIO_CLIENT_FLAG;
2588 smalloc_pool_size = atoi(optarg);
2589 smalloc_pool_size <<= 10;
2593 log_err("fio: --latency-log is deprecated. Use per-job latency log options.\n");
2598 write_bw_log = true;
2603 if (f_out && f_out != stdout)
2606 tmp = fopen(optarg, "w+");
2608 log_err("fio: output file open error: %s\n", strerror(errno));
2613 f_err = f_out = tmp;
2617 output_format = FIO_OUTPUT_TERSE;
2620 if (parse_output_format(optarg)) {
2621 log_err("fio: failed parsing output-format\n");
2628 output_format |= FIO_OUTPUT_TERSE;
2640 fio_show_option_help(optarg);
2647 exit_val = fio_show_ioengine_help(optarg);
2653 dump_cmdline = true;
2661 log_info("%s\n", fio_version_string);
2666 terse_version = atoi(optarg);
2667 if (!(terse_version >= 2 && terse_version <= 5)) {
2668 log_err("fio: bad terse version format\n");
2674 if (!strcmp("always", optarg))
2675 eta_print = FIO_ETA_ALWAYS;
2676 else if (!strcmp("never", optarg))
2677 eta_print = FIO_ETA_NEVER;
2682 if (check_str_time(optarg, &t, 1)) {
2683 log_err("fio: failed parsing eta time %s\n", optarg);
2688 eta_new_line = t / 1000;
2689 if (!eta_new_line) {
2690 log_err("fio: eta new line time too short\n");
2699 if (check_str_time(optarg, &t, 1)) {
2700 log_err("fio: failed parsing eta interval %s\n", optarg);
2705 eta_interval_msec = t / 1000;
2706 if (eta_interval_msec < DISK_UTIL_MSEC) {
2707 log_err("fio: eta interval time too short (%umsec min)\n", DISK_UTIL_MSEC);
2714 if (set_debug(optarg))
2724 if (!strcmp(optarg, "global")) {
2725 log_err("fio: can't use global as only "
2731 new_size = (nr_job_sections + 1) * sizeof(char *);
2732 job_sections = realloc(job_sections, new_size);
2733 job_sections[nr_job_sections] = strdup(optarg);
2739 exit_val = iolog_file_inflate(optarg);
2748 exec_profile = strdup(optarg);
2750 case FIO_GETOPT_JOB: {
2751 const char *opt = l_opts[lidx].name;
2754 if (!strncmp(opt, "name", 4) && td) {
2755 ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
2762 int is_section = !strncmp(opt, "name", 4);
2765 if (!is_section || !strncmp(val, "global", 6))
2768 if (is_section && skip_this_section(val))
2771 td = get_new_job(global, &def_thread, true, NULL);
2772 if (!td || ioengine_load(td)) {
2781 fio_options_set_ioengine_opts(l_opts, td);
2784 if ((!val || !strlen(val)) &&
2785 l_opts[lidx].has_arg == required_argument) {
2786 log_err("fio: option %s requires an argument\n", opt);
2789 ret = fio_cmd_option_parse(td, opt, val);
2800 if (!ret && !strcmp(opt, "ioengine")) {
2801 if (ioengine_load(td)) {
2808 fio_options_set_ioengine_opts(l_opts, td);
2812 case FIO_GETOPT_IOENGINE: {
2813 const char *opt = l_opts[lidx].name;
2819 ret = fio_cmd_ioengine_option_parse(td, opt, val);
2835 /* we don't track/need this anymore, ignore it */
2839 #ifndef CONFIG_NO_SHM
2841 log_err("fio: can't be both client and server\n");
2847 fio_server_set_arg(optarg);
2851 log_err("fio: client/server requires SHM support\n");
2859 fio_server_internal_set(optarg);
2865 pid_file = strdup(optarg);
2868 if ((ret = fio_idle_prof_parse_opt(optarg))) {
2869 /* exit on error and calibration only */
2879 log_err("fio: can't be both client and server\n");
2884 /* if --client parameter contains a pathname */
2885 if (0 == access(optarg, R_OK)) {
2886 /* file contains a list of host addrs or names */
2887 char hostaddr[PATH_MAX] = {0};
2889 FILE * hostf = fopen(optarg, "r");
2891 log_err("fio: could not open client list file %s for read\n", optarg);
2896 sprintf(formatstr, "%%%ds", PATH_MAX - 1);
2898 * read at most PATH_MAX-1 chars from each
2899 * record in this file
2901 while (fscanf(hostf, formatstr, hostaddr) == 1) {
2902 /* expect EVERY host in file to be valid */
2903 if (fio_client_add(&fio_client_ops, hostaddr, &cur_client)) {
2904 log_err("fio: failed adding client %s from file %s\n", hostaddr, optarg);
2911 break; /* no possibility of job file for "this client only" */
2913 if (fio_client_add(&fio_client_ops, optarg, &cur_client)) {
2914 log_err("fio: failed adding client %s\n", optarg);
2920 * If the next argument exists and isn't an option,
2921 * assume it's a job file for this client only.
2923 while (optind < argc) {
2924 if (!strncmp(argv[optind], "--", 2) ||
2925 !strncmp(argv[optind], "-", 1))
2928 if (fio_client_add_ini_file(cur_client, argv[optind], false))
2935 if (fio_client_add_ini_file(cur_client, optarg, true)) {
2943 exit_val = fio_monotonic_clocktest(1);
2948 exit_val = fio_crctest(optarg);
2953 exit_val = fio_memcpy_test(optarg);
2958 if (check_str_time(optarg, &val, 1)) {
2959 log_err("fio: failed parsing time %s\n", optarg);
2965 log_err("fio: status interval too small\n");
2969 status_interval = val / 1000;
2975 trigger_file = strdup(optarg);
2980 trigger_cmd = strdup(optarg);
2983 if (trigger_remote_cmd)
2984 free(trigger_remote_cmd);
2985 trigger_remote_cmd = strdup(optarg);
2990 aux_path = strdup(optarg);
2993 if (check_str_time(optarg, &trigger_timeout, 1)) {
2994 log_err("fio: failed parsing time %s\n", optarg);
2998 trigger_timeout /= 1000000;
3003 merge_blktrace_only = true;
3006 log_err("%s: unrecognized option '%s'\n", argv[0],
3008 show_closest_option(argv[optind - 1]);
3019 if (do_exit && !(is_backend || nr_clients))
3022 if (nr_clients && fio_clients_connect())
3025 if (is_backend && backend)
3026 return fio_start_server(pid_file);
3032 ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
3038 while (!ret && optind < argc) {
3040 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
3041 ini_file[ini_idx - 1] = strdup(argv[optind]);
3049 int fio_init_options(void)
3054 fio_options_fill_optstring();
3055 fio_options_dup_and_init(l_opts);
3059 if (fill_def_thread())
3065 extern int fio_check_options(struct thread_options *);
3067 int parse_options(int argc, char *argv[])
3069 const int type = FIO_CLIENT_TYPE_CLI;
3072 if (fio_init_options())
3074 if (fio_test_cconv(&def_thread.o))
3075 log_err("fio: failed internal cconv test\n");
3077 job_files = parse_cmd_line(argc, argv, type);
3079 if (job_files > 0) {
3080 for (i = 0; i < job_files; i++) {
3081 if (i && fill_def_thread())
3084 if (fio_clients_send_ini(ini_file[i]))
3087 } else if (!is_backend) {
3088 if (parse_jobs_ini(ini_file[i], 0, i, type))
3093 } else if (nr_clients) {
3094 if (fill_def_thread())
3096 if (fio_clients_send_ini(NULL))
3101 fio_options_free(&def_thread);
3102 filesetup_mem_free();
3104 if (!thread_number) {
3109 if (is_backend || nr_clients)
3114 log_err("No job(s) defined\n\n");
3119 if (output_format & FIO_OUTPUT_NORMAL)
3120 log_info("%s\n", fio_version_string);
3125 void options_default_fill(struct thread_options *o)
3127 memcpy(o, &def_thread.o, sizeof(*o));
3130 struct thread_data *get_global_options(void)